kernel_bypass_stack

Setup for Kernel Bypass

• Kernel network stack can be bypassed using an alternate stack of of software like Netmap DPDK and user space stacks like mTCP

Netmap Setup

• Netmap uses vale as the software switch. Follow steps (1-4) from here

• In netmap setup, mTCP requires multi-queue in vale.

• Copy the netmap_vale.c to netmap/sys/dev/netmap/ and compile netmap again. We have tested netmap with igb driver.

• Create interface for each VM with no.of of queues=no of cores of VM.

• ./vale−ctl −n viX −C 2048,2048,Y,Y

  • where Y i s the number o f cores of the VM, X will be a number ( 1 , 2 , 3 .. )

• ./vale−ctl −a vale 0:viX

  • ifconfig viX hw ether 00:aa:bb:cc:dd:0Z //MAC address of the VM

• change ifname in XML to vale0:viX

• Install mTCP in VM as given in the mTCP github page

DPDK Setup:

• We could not get mTCP to work on VMs on DPDK as we were unable to expose multiqueue to the VM. Follow the steps for the DPDK + mTCP setup on host machine:

Using mTCP:

• Make sure to configure and make mTCP whenever you want to change #cores of your VM
  • ./configure --enable-netmap CFLAGS="-DMAX_CPUS=32" if #cores = 32
    • sudo make
  • mTCP configuration needs a server.conf file. You can find an example here
  • Changes required in server.conf
    • port=<network_interface_name>
    • num_cores=<number_of_cores_of_your_VM>
    • For mTCP over DPDK, comment line 6, uncomment line 7
  • Give the path to your mTCP folder in CMakeLists.txt on line 55, 56
  • For mTCP over netmap stack, set the netmap path on line 88 of CMakeLists.txt(../../CMakeLists.txt)
  • Use proper mTCP paths in your VNF Makefile
  • You need not recompile and install libvnf every time you change #cores of VM, but you should recompile your VNF code